How did Lockdown changed crime patterns in certain areas of Greater London?

Introduction

In december of 2019 the first case of a disease involving a 'strange pneumonia' (named later SARS-CoV-2) was detected in Wuhan, China. Though there were some worried voices about the detection of a new unknown disease, comparisons to past epidemics such as the H1N1 and the SARS-MERS brought comfort to leaders and were used to soothe the public opinion. As time passed, scientists started to learn more about this new disease and due to its high transmisibility it soon became clear that there was no point of comparison between the new virus and the ones before it.

By march 2020 the World Health Organization raised the alarms and declared a global pandemic. It became clear that the virus had spread faster and farther than anyone could have imagined and it was evident that decisions to put a brake on the spread were being taken too late. Leaders in Europe realized the magnitude of what they were facing and promptly decided to impose strict lockdown measures even if that meant shutting down the global economy.

caption

After looking at the advances of the virus in it's neighboring countries, the government of UK also decided to act. The 20th of march it was announced that all public venues had to close. This measure applied to pubs, restaurants, gyms, nightclubs, theatres and cinemas. Schools and universities were also included. Three days later it was stated on a television broadcast that a general order to stay at home was inplace and that restrictions on the freedom of movement were going to be enforced by the law. British residents were adviced to stay home throughout this period except for essential purchases, essential work travel (if possible remote work had to be done), medical needs, and providing care for others. These restrictions were eased during late may and june when students were allowed to return to schools and some non essential retailers as well as public venues were allowed to start opening again.

The introduction of the lockdown had an immense effect on London. The center of the city that used to be full of tourists, public venues, social life and a vast amount of commuters had now turned to a rather lonesome place. As the inflow of tourists was severely hindered, non essential retail shops closed and commuters started to work remotely several changes in the city started to manifest.

In this notebook I present an exploratory analysis based on data at the crime event level to answer whether there was any change in the geographic distribution of crime in the greater City of London and in it's composition when Pre Lockdown, Lockdown and Post Lockdown times are compared.

Description of the data

The analysis is based on several sources of data. The main dataset is obtained through the custom download tool provided in the Police Data web page it provides street-level crime and outcomes, broken down by police force and 2011 lower layer super output area (LSOA). We consider the crime data involving both the City Of London Police as well as the Metropolitan Police Service. The period chosen spans from January to September of 2020, based on the rigor of the lockdown measures imposed I categorize the first quarter of the year as Pre Lockdown times, the second quarter as Lockdown and the third one as Post Lockdown.

The information is complemented through geographical data for the LSOA, MSOA (middle super output area) and the London Boroughs. This data is taken from the London Data Store Statistical GIS Boundaries. It contains several shape files that delimit the boundaries of LSOAs, MSOAs and Boroughs in the greater London Area.

Finally, I also consider data obtained through the use of the Police API regarding the number of stop and search procedures carried out by the two police forces mentioned above during the first three quarters of 2020.

All codes used to obtain the main dataset and the API information are included in the notebook Dataset.ipynb. The API data obtained consists of the counts of stop (and possibly search) procedures made on the street by police officers in each of the MSOA's composing the Greater London Area by different age groups. The requests are made through the requests module in Python and results are directly parsed from the response JSON.

Research Question

I am interested in answering the following questions:

An interesting topic I want to address in the last part is to characterize if there exist any associations between types of crimes that ocurred in certain geographical aggregations and the Pre Lockdown, Lockdown and Post Lockdown periods.

Data Cleaning

After gathering the data from the different monthly files obtained by each police force, I carried out some brief data cleaning steps.

Both the id and the Unnamed:0 columns have to be dropped as they don't contain useful information (were mostly used as identifiers in the raw data). Also some renaming of the columns is made to make working with them easier.

The raw data only has missing values in the Last Outcome Category variable, however this is merely used to denote that there was no outcome regarding the crime. It turns out that the missing value is in fact only used to denote the Category of the Crime Anti-social behaviour. So we can just complete it.

I also generate the variable that is going to be used to identify the three periods of interest Pre Lockdown, Lockdown and Post Lockdown. All text variables that represent Categories are reconverted to categorical. The Covid variable is reordered as in the real ocurrence of events and the categories of the variable related to the police force involved Location are recoded. Note that the Location name is rather an identifier of the police force.

Finally, Crimes with counts below 10.000 are dumped into the Other crime category and results are saved in a pickle file.

Data Analysis

Introduction

We start by loading in the already cleaned data from the section above. First thing to notice is that we have 6245 distinct LSOA's and all of these appear in the Metropolitan Service force records.

The so called location variable doesn't represent location but rather the police force that attended the crime. There is an overlap in certain LSOA's for which both the Metropolitan Service and the City of London Police can be in charge of the criminal investigation. The overlap is mainly in the centre of the city (what Londoners know as Zone 1).

In the already cleaned dataset we have 13 different categories of crime types and 15 different outcomes related with these categories. I decided not to lump the low counts in the outcomes variable since the analysis is not going to focus that much in this variable but rather in the types of crime and its comparisons through the quarters of the year.

Evolution of aggregated counts by Police force through the year

To start the analysis its important to consider how the aggregated counts of crime evolved through the year. I present the dynamics of these counts separated by police force involved and considering the aggregate. Data are aggregated for all LSOA's.

Several patterns emerge.

Evolution of aggregated counts by crime type through the year

In order to get a better sense of the changes in crime counts by month I exclude the category of Anti-social behaviour because it generates a large imbalance in the counts. Later I show a barplot in which this category is included but using a logarithmic transformation.

All of these facts are easily seen in the figure below where I plot total crime counts using a logarithmic transformation so as to take into account the notorious difference of counts by type.

Distribution of crime counts by LSOA

While it is true that there are several interesting behaviours in the aggregated data for the Greater London Area there's also a lot of variation in the data at the LSOA level. In this section I explore a little more the results for this geographical disaggregation.

Results are shown using a log transformation over the counts due to the highly skewed distribution of total crimes by location.

The highly skewed distribution observed for the observations regarding the City Of London force (even in logarithm) reflects the fact that most of the crime that this force responds to is located in a couple of LSOA's. By checking the data we note that most of the crime counts in this case happen in the LSOA's City of London 001F and City of London 001G. The first one includes tube stations such as Bank and Liverpool Street while the second is closer to the BlackFriars station. Except for these two aggregates the other locations have a relatively low count of crime cases.

The aggregate distribution closely mirrors the one of the Metropolitan Service this happens again because the high proportion of crime cases that this force has to respond to. The aggregate distribution is bimodal suggesting two types of LSOA's some with very low crime counts and others with a relatively more moderate number of ocurrences.

Overall the second mode of the distribution shifted when comparing the Pre Lockdown, Lockdown and Post Lockdown scenarios. It seems that all in all crime in LSOA's with a low to moderate number of ocurrences increased during the Lockdown and increased further in Post Lockdown when comparing with the results of the first quarter. This fact is probably related with the huge increase in Anti-social behaviour crimes ocurred during the first nine months of 2020.

The table above confirms that there was indeed a shift in the average crime cases when comparing Lockdown and Pre Lockdown with a very important decrease in variability. The maximum number of crime cases in a single LSOA fell from 2271 to only 625 during the Lockdown period. However, once the restrictions were lifted on average crime cases decreased but variability increased again. The maximum number of crime cases in LSOAs doubled from the Lockdown period to the Post Lockdown.

Crime counts for most dangerous LSOA's

The apparent shift that ocurred in the crime distribution during Lockdown along with the decrease in variability merits some further investigation. The figure below shows how crime counts changed in the three periods studied in the 20 LSOA's with the highest number of crime cases during the first nine months of 2020.

We note that overall crime cases fell in each of these LSOA's the decrease was sharper in places close to the center of the City and those that had a high number of cases in Pre Lockdown. Namely, Westminster, Camden and Hackney. This decrease explains a part of the reduction in total variability observed. It is also important to mention that during the Post Lockdown period the crime cases in this LSOA's increased again and reached almost Pre Lockdown levels.

Distribution of crime counts by type

Several conclusions can be obtained from the boxen plots for the distribution of crimes by LSOA.

Criminal Outcomes

The figure below illustrates outcome categories for the different crime events analyzed in the sections above. We note that regarding the categories there doesn't seem to be any huge difference in the results when facetting by Pre, Post and regular Lockdown. Perhaps the only interesting difference is that less crimes were further investigated during the Lockdown period. Nevertheless, this difference just stems from the fact that more Anti-social behaviour crimes ocurred during this period.

We note that no crimes that ocurred during Pre Lockdown times are still under investigation. However, there are lot of cases that ocurred during this period that have an unavailable court result.

Maps by MSOA-Borough

In this section I research more on the geographical distribution of crime dynamics by plotting maps at both the MSOA and the Borough level.

The maps are colored at the MSOA level by the change in the logarithm of aggregate crime counts (approximate variation in cases). The left map compares the change between the Lockdown and Pre Lockdown levels while the second shows the change from the Post Lockdown levels to Lockdown.

The image shows how crime counts sharply fell in the center of the city mostly in Westminster area and the City of London while they started increasing in the outskirts. Some of the most affected MSOA's by this crime absorption phenomena were Hillingdon 05, Merton 025, Bromley 021 and Havering 003.

When comparing the Post Lockdown and the Lockdown levels we note that crime increased again in the center of the city as well as in MSOA's near it. Westminster and the City of London had remarkable increases in crime counts. The Kensigton and Chelsea area also had a strong increase in criminal activity.

Looking at the maps at the Borough level we note one further regularity. While the Lockdown created a strong outflow of criminal activity from the city center to almost all other Boroughs, this dynamic reverted when the Lockdown was lifted. The crime rates for the Boroughs of Kensington and Chelsea and Southwark had increases in both periods. This suggests that the decrease in criminal activity in Zone 1 wasn't completely mirrored in Zone 2 locations. Hence this Zones ended up being affected since they had to face criminal case increases in both periods.

It would be interesting to make some similar maps to note if the conclusion changes when we condition on the type of Crime. Its highly probable that some of the conclusions are mainly driven by the Anti-social behaviour crime given its relatively high frequency.

Has the relation between the number of crimes and searches changed?

In this section I use the search event data to check if there was any change in the number of frisks the police did with respect to the total number of crimes during the Lockdown period.

The figure above shows that the relation between frisks and crime cases is linear in the log-log scale. Its important to mention that there doesn't seem to be any evidence that the relation changed when comparing the three periods. Rater it appears that it remained relatively stable.

Data modelling

In this section I intend to use some matrix clustering procedure and a factorial method to gain further insight on how to group similar LSOA's according to their crime counts in order to learn more about the underlying structures in crime composition.

Matrix Clustering

I took the counts for each crime in the respective LSOA's during the PreLockdown period and calculated a similarity matrix based on this information. I ordered the similarity matrix based on the fiedler eigenvector and created clusters by thresholding at a given value of similarity. The formed groups have the property that the similarity between the first member and the last is above the chosen threshold.

I made these cluster assignments for several threshold values and ultimately chose the best one based on the silhouette value of the assignment.

Given the results of the silhouette for the different thresholds I took 0.45 as the cut point for the initial cluster assignments.

Based on the cluster assignments obtained through the chosen threshold, we see that the procedure separates a group of MSOA's in which crime counts dropped significantly during the Lockdown and another in which counts increased sharply during this period. The first group includes several MSOA's located in the center of the city. The characterizing trait for this group is that crime fell during Lockdown and rose in the Post Lockdown period. However, the increase wasn't enough to get the crime counts to Pre Lockdown levels.

For the second group we note that it is composed of MSOA's in which crime counts increased sharply during the Lockdown period. It seems that for most of the MSOA's the increase in crime counts was temporary and the criminal activity reverted to Pre Lockdown numbers as soon as the Lockdown ended. However, there are certain excepctions to this pattern namely, Bexley 025, Croydon 035, Merton 002, among others.

The result is interesting because it was obtained based only on Pre Lockdown crime counts. What this means is that the crime structure between this two groups was at the onset very different. It is easy to see (by checking the original data) that in fact the MSOA's in group 2 are characterized by a high count of crimes related to theft, this includes Burglary, Theft from the person and other types of theft which were precisely the crimes that fell the most during the Lockdown period precisely due to the stay at home orders.

On the other hand group 0 LSOA's are characterized by high crime counts of Vehicle Crimes and Violence and sexual offences the latter is one of the crimes that increased the most during the first three quarters of the year.

Multiple Correspondence Analysis

In this case since I am interested in the relation of a categorical variable (namely epoch, defined as Pre Lockdown, Lockdown and Post Lockdown) to a set of continuous variables I can't use traditional dimension reduction methods because I would lose the variable I am most interested in. Hence, to face this problem I decided to bin the continuous variables and carried out a Multiple Correspondence Analysis.

The first figure above shows the projection of the variables in the factorial space. We note that the first factorial axis separates low, medium and high counts of the different crimes. The left part of the axis is related to low counts in any of the categories while the right part is related to high ones. The second factorial axis creates a separation between Pre Lockdown times (coded as Low intensity) and Lockdown|PostLockdown (coded as Medium and High intensity respectively). We note that the separation obtained in the crime variables is better than the one obtained in the periods.

Conclusion

Overall we can make the following characterization of the crime composition in the three periods considered. During the Pre Lockdown period there was an important contribution of geographical aggregations located in the city center where crimes related to general forms of theft ocurred frequently. The introduction of the lockdown disrupted the inflow of people in the city center (both conmuters and tourists) this was apparently associated with a notorious decrease in theft cases and at the macrolevel a general reduction in criminal activity in Boroughs such as the City Of London and Westminster.

During the Lockdown the sharp decrease in theft was reflected by a less amount of cases being handled by the City Of London police force and a higher proportion received by the Metropolitan Service Force. There was a very strong increase in cases of Anti social behaviour and a relatively moderate one in Drug crimes. The ocurrence of these crimes was more related to peripheral areas rather than the city center.

After the end of the Lockdown there was a new surge of criminal activity in the city center and notable recover of counts related with general forms of theft in this location. Probably related with the increase in the flow of conmuters and general public. Drug crimes and Anti Social behaviour related ocurrences also fell in the Post Lockdown though the latter remained above its historical levels (when comparing with the first quarter of the year). However, the Post Lockdown period also showed an alarming increase in crimes related with Violence and Sexual offences which should be investigated further.

Overall this crime composition changes seem to have had worsened the situation in some Zone 2 geographical aggreations that experiment increases both when comparing the Lockdown period with the Pre Lockdown and the Post Lockdown with the former one. There was a steady increase in Bicycle Theft cases since the introduction of the Lockdown this trend consolitade during the last quarter of the year. Finally, there wasn't any apparent change in the number of search procedures carried by the police during the year with respect to the crime counts.